OpenAI Astra Hits Critical Cyber Threshold

The real story is not another benchmark jump—it is the point where deployment controls become part of the product.

• –Astra achieved 100% on ExploitBench and reportedly discovered two zero-day vulnerabilities in a newer internal evaluation.
• –Its results reflect Daybreak Blue access, not the default production configuration, so independent verification remains essential.
• –OpenAI says Astra is more capable and token-efficient than GPT-5.6 Sol for vulnerability identification and exploit development.
• –Restricted access, chain-of-thought monitoring, refusals, and automatic interruption will likely create friction for legitimate defensive security work.
• –The model’s eventual system card and external testing will determine whether OpenAI’s “Critical” designation is a reproducible capability claim or primarily an internal risk label.